The Open Secure AI Alliance, now more than 120 organizations strong, opened a Request for Comments today on Shared AI Findings Exchange (SAFE), a Linux Foundation-hosted framework for collecting and analyzing agentic AI security incidents across the industry. The proposal landed on August 4, 2026, timed to the opening of Black Hat in Las Vegas, and is being drafted by an alliance working group that includes NVIDIA, Cisco, CrowdStrike, Hugging Face, and Red Hat. The pitch is straightforward: turn each agent breach or near miss into shared defensive intelligence rather than a siloed embarrassment.
SAFE would confidentially gather incident reports and near misses, notify affected parties, identify recurring control failures, and publish evidence-based operating recommendations. The model borrows from decades-old aviation and traditional cybersecurity information-sharing regimes, but retrofits them for the specific failure modes of AI agents — prompt injection, tool poisoning, and unauthorized action chains that don't map cleanly onto CVE-style vulnerability disclosure.
Justin Boitano, who leads NVIDIA's enterprise AI efforts, framed the launch as a response to the widening attack surface introduced by autonomous agents. The framing matters because agents are being deployed inside enterprises faster than security teams can audit them, and the industry has no common language for describing what went wrong when one misbehaves.
“Defenders must move now at agent speed to respond rapidly to protect infrastructure and intellectual property — and the best way to do that is together. When trusted ecosystems share threat intelligence openly, collective defense becomes a force multiplier.”— Justin Boitano, NVIDIA
Key facts
- 01The Open Secure AI Alliance now counts more than 120 member organizations, including NVIDIA, Cisco, CrowdStrike, Hugging Face, and Red Hat.
- 02The Linux Foundation opened a Request for Comments on Shared AI Findings Exchange (SAFE), a framework to share agentic AI incidents and near misses.
- 03CrowdStrike's fine-tuned NVIDIA Nemotron Nano model hit 96% accuracy generating investigation queries in Falcon LogScale.
- 04Amazon and Visa joined the alliance today, contributing Strands Agents, Cedar, and the Visa Vulnerability Agentic Harness as open source projects.
- 05The announcement lands as Black Hat begins August 4, 2026 in Las Vegas.
NVIDIA's stack contribution runs end to end. The NVIDIA Labs Object-Oriented Agent (NOOA) research harness on GitHub is meant to make agent behavior testable and auditable. NVIDIA OpenShell acts as a runtime sandbox that restricts what an agent can see and touch. Open-weight model families — Nemotron for agentic AI, Cosmos for physical AI, Isaac GR00T for robotics, BioNeMo for life sciences, and Alpamayo for autonomous vehicles — ship with datasets and training techniques exposed. NeMo Guardrails, NeMo Anonymizer, and NeMo Safe Synthesizer handle policy enforcement and synthetic data, while Garak scans models for jailbreaks and data leaks before deployment.
CrowdStrike disclosed the most concrete performance number in the announcement: internal testing of a fine-tuned NVIDIA Nemotron Nano model hit 96% accuracy generating investigation queries inside Falcon LogScale. The company also published research showing the same specialized reasoning model outperforming much larger general-purpose models on Security Operations Center detection triage, with calibrated confidence scoring that makes autonomous security decisions auditable.
Cisco released two of its Antares small language models for locating known vulnerabilities in codebases, plus Project CodeGuard for embedding secure defaults in AI coding workflows. Cisco DefenseClaw sits on top of NVIDIA OpenShell to govern agentic workforces at runtime.
Amazon and Visa both joined the alliance today. Amazon contributed Strands Agents, an open toolkit for building and evaluating agents in production, and Cedar, an open authorization language for enforcing deterministic boundaries on what agents can do. Visa open-sourced its Vulnerability Agentic Harness for identifying and validating security issues. Okta is building reference implementations for Cross App Access (XAA), a protocol letting agents inside OpenShell sandboxes connect to enterprise apps under identity controls. Palo Alto Networks contributed Agent Guard and Agent Watch from its Idira platform. Red Hat founded asago, which maps organizational governance requirements — including NIST, OWASP, and EU AI Act provisions — to runtime agent permissions with an audit trail from policy to enforcement.
Other contributions round out the stack. Microsoft AI Red Team open-sourced PyRIT for automated red teaming, RAMPART for turning findings into repeatable tests, Clarity for pre-code design review, and Assert for converting requirements into executable safety evaluations. Capital One released VulnHunter for agentic code security. Cloudflare offered a Vulnerability Discovery Harness. Perplexity contributed Numbat, an endpoint agent security suite for macOS, Linux, and Windows. Uber open-sourced components of ADR, its Agentic AI Detection and Response system. Wiz released Atlas, an autonomous vulnerability research engine.
The obvious caveat is that a Request for Comments is not a standard, and voluntary incident-sharing frameworks have a mixed track record. Traditional cyber threat intel exchanges took years to gain traction even with regulatory nudges, and companies remain reluctant to disclose incidents that could expose them to liability or reputational damage. Whether SAFE gains the confidential-reporting protections that made aviation safety reporting work will determine if enterprises actually file. The absence today of OpenAI, Anthropic, and Google from the named contributor list is also notable — the frontier labs whose models power most agentic deployments are not visible in the alliance's founding tier.
The strategic read is that NVIDIA is trying to make agent security an open-source, ecosystem-wide problem rather than a per-vendor feature war, and doing so at a moment when it can shape the defaults. Every hyperscaler and security vendor now shipping agents has an interest in a shared incident taxonomy — without one, the first major agent-driven breach becomes a regulatory event that hits everyone. Getting 120 organizations behind a Linux Foundation framework before that breach happens is a defensive play for the entire agentic AI market, and it costs NVIDIA very little to lead it.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
